Preface

As Agents are used over time, the most common issue is repeating the same mistakes: a command fails twice, the same correction needs to be told to the Agent a third time, and it still doesn’t learn. The problem isn’t the model, but that experience is lost at the end of each session, with no mechanism to consolidate lessons learned for future use. The proactive-self-improving-agent Skill addresses this exact issue: it adds a set of trigger, record, review, and promotion workflows for experience management, enabling the Agent to automatically capture experience and evolve safely during operation.

What is this

This is a community-published Skill, located at @user_96b3f16c/proactive-self-improving-agent, version 1.0.0, categorized under ai-agent. It does not alter the Agent’s inherent reasoning capabilities; instead, it attaches a mechanism to the Agent’s workflow: under specific trigger conditions, it writes experience into Markdown files within the workspace. It controls write frequency through deduplication and logging to avoid noise accumulation. The core problem it solves is that lessons learned generated during task completion are not retained by default, meaning the Agent has to start from scratch when encountering similar tasks again.

Core Features: Seven Trigger Conditions

The Skill’s trigger conditions fall into seven categories, each corresponding to different recording actions:

  1. When a command or operation fails, record it in ERRORS.md.
  2. When corrected by the user (identifying phrases like “not right” or “should be”), record it in LEARNINGS.md.
  3. When the user requires a capability that does not exist, record it in FEATURE_REQUESTS.md.
  4. When an external API or tool errors, record it in ERRORS.md.
  5. When discovering that one’s knowledge is outdated or incorrect, record it in LEARNINGS.md.
  6. When finding a better approach, record it in LEARNINGS.md.
  7. Upon completion of each task, review the entire process; if there is new experience, record it in LEARNINGS.md.

As can be seen, failure-type experiences (1, 4) are uniformly written to ERRORS.md, correction and knowledge update-type experiences (2, 5, 6, 7) go to LEARNINGS.md, and capability gaps are recorded separately in FEATURE_REQUESTS.md. The three files have clear divisions of labor, facilitating subsequent retrieval and auditing.

Deduplication Principles and Logging

Writing is not unconditional. The Skill explicitly requires skipping the write if there is no new experience, or if an existing entry already covers the current experience. This deduplication principle is key to the long-term operation of the entire mechanism—if experience files only grow without reduction and duplicate entries pile up, they will quickly lose their reference value.

Every time an experience is written, a log entry in JSONL format must also be appended to .learnings/CHANGELOG.md simultaneously. JSONL means one JSON object per line; this format facilitates appending via scripts and subsequent parsing, while also leaving a complete timeline of writes.

The Promotion Path for Experience

Experience does not remain in LEARNINGS.md permanently. When a piece of experience reoccurs and reaches 3 instances, it is promoted to more formal, frequently read files such as AGENTS.md, TOOLS.md, or SOUL.md. Experience appearing only once stays in the .learnings/ directory, while experience verified effective multiple times rises to a higher level, forming a screening mechanism from temporary records to long-term rules.

Installation and Activation

The official page (https://api.skillhub.cn/user_96b3f16c/proactive-self-improving-agent and the corresponding skillhub.cn page) was unreachable at the time of writing, so installation commands and specific directory placement could not be verified; no speculative content is provided here. From the Skill’s working method, it can be inferred that it requires the ability to create ERRORS.md, LEARNINGS.md, FEATURE_REQUESTS.md, and the .learnings/ directory within the Agent’s working directory. Refer to the official page for specific activation methods.

Use Cases and Caveats

It is suitable for scenarios where the Agent is used long-term within the same working directory: project development, daily operational scripts, and workflows requiring repeated handling of similar tasks. Experience files are consolidated within the workspace; switching directories means a new set of memory. This is by design, but it also implies that experience will not automatically migrate across directories.

A few points to note:

  1. Experience writing relies on the identification of trigger conditions. User correction triggers are based on phrases like “not right” or “should be”; if the wording during correction is not typical, it may not be captured.
  2. Deduplication and the 3-time promotion rule require accurate execution by the Agent; the effectiveness of the mechanism itself depends on how well the Agent adheres to the rules.
  3. Experience promoted to files like AGENTS.md will remain in effect; it is worth a manual review before promotion to avoid solidifying incidental situations into rules.
  4. This Skill is marked as requires_api_key=false, requiring no additional API key configuration.

Conclusion

The value of this type of Skill lies not in making the Agent smarter, but in preventing it from repeating mistakes: solidifying failures, corrections, and discoveries into files, and controlling scale through deduplication and promotion mechanisms. If your Agent has been running in a project for a while, it is worthwhile to use this to systematically organize scattered experiences.

Official URL: https://api.skillhub.cn/user_96b3f16c/proactive-self-improving-agent